Pictures of gno.land pages: the whole gnoweb page, and the crop that actually matters, the realm's own rendered output.
Every gno.land link shared today previews as a bare text card, and every list of realms in every explorer is a column of monospace paths. This turns both into pictures.
gnoshot capture https://gno.land/r/gov/dao -out /tmp/dao
probe status=200 state=ok freshness=7f0a6b4c2441348b
capture 1.511s matched=realm selector=md-renderer.c-realm-view box=816x1358
master-render.webp 74944 bytes (1632,2714) truncated=false
master-page.webp 150064 bytes (2560,3254) truncated=false
og.webp 16570 bytes
hero@1x.webp 10788 bytes
hero@2x.webp 25644 bytes
thumb@1x.webp 1230 bytes
thumb@2x.webp 3640 bytes
go install github.com/gnoverse/gnoshot@latest
It needs a Chrome or Chromium on the box. It looks in the usual places and in
Playwright's download directory; GNOSHOT_CHROME overrides.
gnoshot serve -root /var/lib/gnoshot -source https://mygnoscan.moul.p2p.team
| route | |
|---|---|
GET /shot?url=&mode=page|render&size=og|hero|thumb&theme=light|dark&dpr=1|2&v= |
the image |
GET /meta?url=&theme= |
the manifest entry: which selector matched, the box, whether it was truncated |
GET /tile?path=&kind=&size=&dpr= |
a generated placeholder, no capture involved |
POST /invalidate |
{"url": "..."}, bearer token, forces a re-probe |
GET /healthz |
queue depth, workers, entries, oldest probe |
/shot never answers 404. A gno.land URL that does not resolve still gets an
image, because the consumer is an <img> and its only other option is the
broken-image glyph.
Two modes. page is the whole gnoweb page, chrome included. render is the
realm's own output, and it is the one worth showing.
The crop is a chain, not a pinned selector, resolved by first candidate with a non-zero bounding box, and which one matched is part of the response:
| matched | selector | what it means |
|---|---|---|
realm |
md-renderer.c-realm-view |
a realm render, the primary capture |
readme |
md-renderer.c-readme-view |
a documented directory or pure package, still a good crop |
directory |
article.b-directory |
an undocumented directory listing |
status |
(nothing matched) | an error, a 404, or a package not yet enabled |
Verified against live gnoweb on gnoland-1, 2026-09-22, and reproducible with
make selectors:
HTTP STATE SELECTOR BOX URL
200 realm md-renderer.c-realm-view 816x1358 https://gno.land/r/gov/dao
200 readme md-renderer.c-readme-view 1200x384 https://gno.land/r/moul/config/v0
200 directory article.b-directory 1200x185 https://gno.land/r/moul/config
200 directory article.b-directory 1200x185 https://gno.land/p/moul/txlink
200 realm md-renderer.c-realm-view 816x1154 https://gno.land/u/moul
404 inert - - .../pixelgnomes
404 not_found - - https://gno.land/r/does/not/exist12345
Three things that look like details and are not:
- The class has to be pinned. A bare
md-rendereralso matches$sourceand$help, neither of which is a realm render and both of which are thousands of pixels tall. - Non-zero area, not a null check.
md-renderer.c-realm-viewhas appeared on/u/<user>with width and height 0. Asking Chrome to photograph a zero-area node does not fail, it never returns, and one such URL in the queue holds a worker for the whole deadline. - 404 is two different things. A package that has been submitted and is not yet enabled answers 404 with the same full-screen box as one that does not exist. They are told apart by the body text, because the status code cannot, and the first gets a 5-minute negative TTL because it flips to a real render with no request from us.
The crop is the matched element's own bounding box, which means anything not inside that box is not in the picture. A render captured raw has its first heading glued to the pixel at 0,0 and its last line running off the edge.
So the element is padded before it is measured, and the breathing room lands inside the crop rather than around it.
Padding only, never a width. Widening the element to fill a 16:9 frame is the obvious next step and it is wrong: gnoweb lays the realm view beside an "On this page" sidebar, so a wider box does not gain empty space, it grows over the neighbour and photographs half a table of contents down the right edge.
Changing any of this bumps RenderVersion, which is part of the cache key.
The key is otherwise a hash of the gnoweb response, and that answers "has the
page changed" rather than "would we photograph it differently today". Without
the bump, every already-captured realm keeps the old framing forever, because
the page did not move.
Capture once at device scale factor 2, derive everything else. Re-rendering per size spends a browser round trip to produce pixels a resize already has.
| rung | candidates | for |
|---|---|---|
og |
1200x630 | og:image and twitter:image; crawlers fetch one fixed size |
hero |
640x360, 1280x720 | a detail-page hero, srcset 1x/2x |
thumb |
160x90, 320x180 | list rows and cards, srcset 1x/2x |
Master in lossless WebP, derivatives in lossy. The rule inverts below
native resolution: on a native-resolution render lossless is 1.9x smaller than
q80, and on the downscaled og rung it is 2.9x larger, because resampling
turns crisp glyph edges into the gradients a DCT codec is good at.
There is no avatar rung. A square crop from the top of a render 816 CSS px
wide is an 8.5x downscale, which lands 16 px body text under 2 px tall: not a
picture of a realm, an illegible fragment of its first heading. Use an identicon
for a square identity mark; it is honestly synthetic rather than dishonestly
photographic.
Never a broken image and never a blank one. A blank box reads as "the site is broken", not "this one has no render yet", and the reader this exists for does not know the difference.
Six states, each with a short label for a dense table and a plain line for
everyone else. The plain line belongs in the alt at every size, because a
screen reader has no size.
| kind | label | plain line |
|---|---|---|
parked |
not enabled yet | Published, but still waiting for approval before anyone can use it. |
pure |
code library | Building blocks other programs use. It has no page of its own. |
no-render |
no page to show | This program works. It just does not display anything. |
directory |
folder | A folder, not a program. The things inside it have their own pages. |
none |
no preview | No picture available for this one. |
error |
unavailable | Could not load a picture of this one just now. |
The colour is seeded from sha256(path), so a package gets the same tile on
every surface and across restarts, and the tile is cacheable forever.
The cache key is the sha256 of the gnoweb response body, not the URL and not
a TTL. A realm's render is a function of chain state: it changes when nobody
redeploys anything, and it stays put through a hundred page views. The body hash
is exact, and it covers the :path arguments and the gnoweb build id for free.
The probe that computes it is one HTTP GET, ~45 ms, against ~500 ms for
vm/qrender, which would not see either of those anyway.
Serving is stale-while-revalidate: a cached picture goes out immediately and the re-probe happens behind it. A reader never waits on Chrome.
Never immutable, and that is deliberate. A caller pins v= to something
about the page, typically a block height. The bytes also depend on the recipe
here, which v= cannot see: change how a render is framed and every returning
visitor keeps the old picture until their v= moves, which for a genesis
package nobody has ever called is never. max-age plus
stale-while-revalidate keeps the paint instant and lets the ETag answer 304
behind it.
A transient upstream error never overwrites a good picture. The realm was fine a minute ago and will be again, and an error tile in place of a slightly old screenshot is the worse answer.
?url= is restricted to an allowlist. An unbounded one is an open proxy and a
way to spend this box's CPU on headless Chrome. POST /invalidate needs a
bearer token and is refused outright when none is configured.
There are two lists, and the split says what kind of page a host serves:
| flag | for | how it is photographed |
|---|---|---|
-allow |
gnoweb hosts (default gno.land and the testnets) |
through the selector chain, cropped to the realm's own render |
-allow-site |
anything else, empty by default | whole page, no selector chain |
The second exists because "no selector in the chain matched" means two
different things depending on where you are. On gnoweb it means an error page,
which is what matched=status records and why a placeholder tile is served
instead of a screenshot. On somebody else's app it means nothing at all: the
chain is gnoweb's own class names. A site host is recorded as matched=site,
is never demoted to a status tile for failing to look like gnoweb, and the
ladder is derived from its page master, so size=thumb and size=og answer
normally.
Keeping the two apart is deliberate: the operator states which kind a host is, rather than the service guessing from what it found. A host named in both is a configuration mistake and gnoweb wins, because the selector chain produces a better picture wherever it applies.
Chrome is the part everyone expects to be slow and it is not. Measured on the
deployment box, 2026-09-22, /r/gov/dao render only: 4.6 s end to end, of
which the capture is 0.8 s and a browser launch is ~0.5 s and happens once per
worker. The rest is WebP.
That is why the sweep captures the render and not the page. The page master is 2560x3254 against the render's 1632x2714, nothing derives a rung from it, and the pixel count is the whole cost. A page master is captured when somebody actually asks for one, and one already on disk survives a render-only refresh.
Two things that are not levers, both measured rather than assumed:
- Encoder effort. Methods 1 through 6 produce byte-identical output on the same image, within 0.93 to 1.23 s. Only method 0 differs, at 7.7x the bytes.
- Lossy for the page master. 261,280 bytes at q90 against 150,064 lossless on the same page. The rule only inverts below native resolution.
RAM is the constraint rather than CPU: roughly 1.3 GB per worker. Storage is ~300 KB per path with both masters, ~180 KB with the render alone.
Two workers on a box that already runs something else is the right size.
gnoshot serve run the service
gnoshot capture <url> [-out DIR] [-mode ...] one URL, the full ladder, no service
gnoshot resolve <url>... which selector matches, and the box
gnoshot ladder <master.png> derivatives only, no browser
gnoshot tile <path> [-kind ...] one placeholder
gnoshot sweep [-service URL] enqueue every known path
capture and the worker run the identical code path, so a picture that is wrong
in production is reproducible with one command and nothing running.